If this is useful, a ❤️ helps others find it.

I run both in production. Here's the real comparison — not theoretical, from actual use building developer tools.

Side by side

Quality in practice

Simple tasks (summarize, classify, format):
Local 7B model = Gemini Flash. Indistinguishable for basic tasks.

Complex reasoning (debug a crash, trace causality, explain why):
Gemini wins clearly. A local 7B model struggles with multi-step reasoning chains.

Code completion (autocomplete, short snippets):
Local 1.5B model (qwen2.5-coder) is fast enough and good enough. No need to send code to cloud.

When local wins

• You're processing medical records, legal documents, financial data
• Your users are on corporate networks with strict egress policies
• You need zero latency (model already loaded, no network round-trip)
• You're building for offline use

When Gemini wins

• You need the best reasoning quality available
• Your data isn't sensitive
• Your users won't install a 4GB model to try your app
• You're prototyping and want to move fast

The hybrid approach (what I actually do)

Code autocomplete → Local (qwen2.5-coder:1.5b, instant)
Log diagnosis → Gemini API (better reasoning, PII filtered)
PDF processing → Local (privacy-sensitive documents)
General chat → Gemini API (quality matters)

Not either/or. Each tool for the right job.

Hardware reality for local LLMs

On an 8-year-old MacBook Air (8GB RAM, Intel):

• qwen2.5-coder:1.5b → fast, great for autocomplete
• gemma2 (9B) → slow first token (~8s), usable
• llama3 (8B) → similar to gemma2
• Anything 70B → not viable, not enough RAM

Apple Silicon (M-series) runs local LLMs significantly better due to unified memory. If you're on M1/M2/M3, local quality improves substantially.

Hiyoko PDF Vault → https://hiyokoko.gumroad.com/l/HiyokoPDFVault
X → @hiyoyok